xref: /llvm-project/llvm/lib/Transforms/Scalar/DivRemPairs.cpp (revision 31b6093434e25bc6cf20ed9ee479557ebbe938f3)
1 //===- DivRemPairs.cpp - Hoist/[dr]ecompose division and remainder --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass hoists and/or decomposes/recomposes integer division and remainder
10 // instructions to enable CFG improvements and better codegen.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Scalar/DivRemPairs.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/MapVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/GlobalsModRef.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/PatternMatch.h"
24 #include "llvm/InitializePasses.h"
25 #include "llvm/Pass.h"
26 #include "llvm/Support/DebugCounter.h"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
29 #include <optional>
30 
31 using namespace llvm;
32 using namespace llvm::PatternMatch;
33 
34 #define DEBUG_TYPE "div-rem-pairs"
35 STATISTIC(NumPairs, "Number of div/rem pairs");
36 STATISTIC(NumRecomposed, "Number of instructions recomposed");
37 STATISTIC(NumHoisted, "Number of instructions hoisted");
38 STATISTIC(NumDecomposed, "Number of instructions decomposed");
39 DEBUG_COUNTER(DRPCounter, "div-rem-pairs-transform",
40               "Controls transformations in div-rem-pairs pass");
41 
42 namespace {
43 struct ExpandedMatch {
44   DivRemMapKey Key;
45   Instruction *Value;
46 };
47 } // namespace
48 
49 /// See if we can match: (which is the form we expand into)
50 ///   X - ((X ?/ Y) * Y)
51 /// which is equivalent to:
52 ///   X ?% Y
53 static std::optional<ExpandedMatch> matchExpandedRem(Instruction &I) {
54   Value *Dividend, *XroundedDownToMultipleOfY;
55   if (!match(&I, m_Sub(m_Value(Dividend), m_Value(XroundedDownToMultipleOfY))))
56     return std::nullopt;
57 
58   Value *Divisor;
59   Instruction *Div;
60   // Look for  ((X / Y) * Y)
61   if (!match(
62           XroundedDownToMultipleOfY,
63           m_c_Mul(m_CombineAnd(m_IDiv(m_Specific(Dividend), m_Value(Divisor)),
64                                m_Instruction(Div)),
65                   m_Deferred(Divisor))))
66     return std::nullopt;
67 
68   ExpandedMatch M;
69   M.Key.SignedOp = Div->getOpcode() == Instruction::SDiv;
70   M.Key.Dividend = Dividend;
71   M.Key.Divisor = Divisor;
72   M.Value = &I;
73   return M;
74 }
75 
76 namespace {
77 /// A thin wrapper to store two values that we matched as div-rem pair.
78 /// We want this extra indirection to avoid dealing with RAUW'ing the map keys.
79 struct DivRemPairWorklistEntry {
80   /// The actual udiv/sdiv instruction. Source of truth.
81   AssertingVH<Instruction> DivInst;
82 
83   /// The instruction that we have matched as a remainder instruction.
84   /// Should only be used as Value, don't introspect it.
85   AssertingVH<Instruction> RemInst;
86 
87   DivRemPairWorklistEntry(Instruction *DivInst_, Instruction *RemInst_)
88       : DivInst(DivInst_), RemInst(RemInst_) {
89     assert((DivInst->getOpcode() == Instruction::UDiv ||
90             DivInst->getOpcode() == Instruction::SDiv) &&
91            "Not a division.");
92     assert(DivInst->getType() == RemInst->getType() && "Types should match.");
93     // We can't check anything else about remainder instruction,
94     // it's not strictly required to be a urem/srem.
95   }
96 
97   /// The type for this pair, identical for both the div and rem.
98   Type *getType() const { return DivInst->getType(); }
99 
100   /// Is this pair signed or unsigned?
101   bool isSigned() const { return DivInst->getOpcode() == Instruction::SDiv; }
102 
103   /// In this pair, what are the divident and divisor?
104   Value *getDividend() const { return DivInst->getOperand(0); }
105   Value *getDivisor() const { return DivInst->getOperand(1); }
106 
107   bool isRemExpanded() const {
108     switch (RemInst->getOpcode()) {
109     case Instruction::SRem:
110     case Instruction::URem:
111       return false; // single 'rem' instruction - unexpanded form.
112     default:
113       return true; // anything else means we have remainder in expanded form.
114     }
115   }
116 };
117 } // namespace
118 using DivRemWorklistTy = SmallVector<DivRemPairWorklistEntry, 4>;
119 
120 /// Find matching pairs of integer div/rem ops (they have the same numerator,
121 /// denominator, and signedness). Place those pairs into a worklist for further
122 /// processing. This indirection is needed because we have to use TrackingVH<>
123 /// because we will be doing RAUW, and if one of the rem instructions we change
124 /// happens to be an input to another div/rem in the maps, we'd have problems.
125 static DivRemWorklistTy getWorklist(Function &F) {
126   // Insert all divide and remainder instructions into maps keyed by their
127   // operands and opcode (signed or unsigned).
128   DenseMap<DivRemMapKey, Instruction *> DivMap;
129   // Use a MapVector for RemMap so that instructions are moved/inserted in a
130   // deterministic order.
131   MapVector<DivRemMapKey, Instruction *> RemMap;
132   for (auto &BB : F) {
133     for (auto &I : BB) {
134       if (I.getOpcode() == Instruction::SDiv)
135         DivMap[DivRemMapKey(true, I.getOperand(0), I.getOperand(1))] = &I;
136       else if (I.getOpcode() == Instruction::UDiv)
137         DivMap[DivRemMapKey(false, I.getOperand(0), I.getOperand(1))] = &I;
138       else if (I.getOpcode() == Instruction::SRem)
139         RemMap[DivRemMapKey(true, I.getOperand(0), I.getOperand(1))] = &I;
140       else if (I.getOpcode() == Instruction::URem)
141         RemMap[DivRemMapKey(false, I.getOperand(0), I.getOperand(1))] = &I;
142       else if (auto Match = matchExpandedRem(I))
143         RemMap[Match->Key] = Match->Value;
144     }
145   }
146 
147   // We'll accumulate the matching pairs of div-rem instructions here.
148   DivRemWorklistTy Worklist;
149 
150   // We can iterate over either map because we are only looking for matched
151   // pairs. Choose remainders for efficiency because they are usually even more
152   // rare than division.
153   for (auto &RemPair : RemMap) {
154     // Find the matching division instruction from the division map.
155     auto It = DivMap.find(RemPair.first);
156     if (It == DivMap.end())
157       continue;
158 
159     // We have a matching pair of div/rem instructions.
160     NumPairs++;
161     Instruction *RemInst = RemPair.second;
162 
163     // Place it in the worklist.
164     Worklist.emplace_back(It->second, RemInst);
165   }
166 
167   return Worklist;
168 }
169 
170 /// Find matching pairs of integer div/rem ops (they have the same numerator,
171 /// denominator, and signedness). If they exist in different basic blocks, bring
172 /// them together by hoisting or replace the common division operation that is
173 /// implicit in the remainder:
174 /// X % Y <--> X - ((X / Y) * Y).
175 ///
176 /// We can largely ignore the normal safety and cost constraints on speculation
177 /// of these ops when we find a matching pair. This is because we are already
178 /// guaranteed that any exceptions and most cost are already incurred by the
179 /// first member of the pair.
180 ///
181 /// Note: This transform could be an oddball enhancement to EarlyCSE, GVN, or
182 /// SimplifyCFG, but it's split off on its own because it's different enough
183 /// that it doesn't quite match the stated objectives of those passes.
184 static bool optimizeDivRem(Function &F, const TargetTransformInfo &TTI,
185                            const DominatorTree &DT) {
186   bool Changed = false;
187 
188   // Get the matching pairs of div-rem instructions. We want this extra
189   // indirection to avoid dealing with having to RAUW the keys of the maps.
190   DivRemWorklistTy Worklist = getWorklist(F);
191 
192   // Process each entry in the worklist.
193   for (DivRemPairWorklistEntry &E : Worklist) {
194     if (!DebugCounter::shouldExecute(DRPCounter))
195       continue;
196 
197     bool HasDivRemOp = TTI.hasDivRemOp(E.getType(), E.isSigned());
198 
199     auto &DivInst = E.DivInst;
200     auto &RemInst = E.RemInst;
201 
202     const bool RemOriginallyWasInExpandedForm = E.isRemExpanded();
203     (void)RemOriginallyWasInExpandedForm; // suppress unused variable warning
204 
205     if (HasDivRemOp && E.isRemExpanded()) {
206       // The target supports div+rem but the rem is expanded.
207       // We should recompose it first.
208       Value *X = E.getDividend();
209       Value *Y = E.getDivisor();
210       Instruction *RealRem = E.isSigned() ? BinaryOperator::CreateSRem(X, Y)
211                                           : BinaryOperator::CreateURem(X, Y);
212       // Note that we place it right next to the original expanded instruction,
213       // and letting further handling to move it if needed.
214       RealRem->setName(RemInst->getName() + ".recomposed");
215       RealRem->insertAfter(RemInst);
216       Instruction *OrigRemInst = RemInst;
217       // Update AssertingVH<> with new instruction so it doesn't assert.
218       RemInst = RealRem;
219       // And replace the original instruction with the new one.
220       OrigRemInst->replaceAllUsesWith(RealRem);
221       OrigRemInst->eraseFromParent();
222       NumRecomposed++;
223       // Note that we have left ((X / Y) * Y) around.
224       // If it had other uses we could rewrite it as X - X % Y
225       Changed = true;
226     }
227 
228     assert((!E.isRemExpanded() || !HasDivRemOp) &&
229            "*If* the target supports div-rem, then by now the RemInst *is* "
230            "Instruction::[US]Rem.");
231 
232     // If the target supports div+rem and the instructions are in the same block
233     // already, there's nothing to do. The backend should handle this. If the
234     // target does not support div+rem, then we will decompose the rem.
235     if (HasDivRemOp && RemInst->getParent() == DivInst->getParent())
236       continue;
237 
238     bool DivDominates = DT.dominates(DivInst, RemInst);
239     if (!DivDominates && !DT.dominates(RemInst, DivInst)) {
240       // We have matching div-rem pair, but they are in two different blocks,
241       // neither of which dominates one another.
242 
243       BasicBlock *PredBB = nullptr;
244       BasicBlock *DivBB = DivInst->getParent();
245       BasicBlock *RemBB = RemInst->getParent();
246 
247       // It's only safe to hoist if every instruction before the Div/Rem in the
248       // basic block is guaranteed to transfer execution.
249       auto IsSafeToHoist = [](Instruction *DivOrRem, BasicBlock *ParentBB) {
250         for (auto I = ParentBB->begin(), E = DivOrRem->getIterator(); I != E;
251              ++I)
252           if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
253             return false;
254 
255         return true;
256       };
257 
258       // Look for something like this
259       // PredBB
260       //   |  \
261       //   |  Rem
262       //   |  /
263       //  Div
264       //
265       // If the Rem block has a single predecessor and successor, and all paths
266       // from PredBB go to either RemBB or DivBB, and execution of RemBB and
267       // DivBB will always reach the Div/Rem, we can hoist Div to PredBB. If
268       // we have a DivRem operation we can also hoist Rem. Otherwise we'll leave
269       // Rem where it is and rewrite it to mul/sub.
270       // FIXME: We could handle more hoisting cases.
271       if (RemBB->getSingleSuccessor() == DivBB)
272         PredBB = RemBB->getUniquePredecessor();
273 
274       if (PredBB && IsSafeToHoist(RemInst, RemBB) &&
275           IsSafeToHoist(DivInst, DivBB) &&
276           all_of(successors(PredBB),
277                  [&](BasicBlock *BB) { return BB == DivBB || BB == RemBB; }) &&
278           all_of(predecessors(DivBB),
279                  [&](BasicBlock *BB) { return BB == RemBB || BB == PredBB; })) {
280         DivDominates = true;
281         DivInst->moveBefore(PredBB->getTerminator());
282         Changed = true;
283         if (HasDivRemOp) {
284           RemInst->moveBefore(PredBB->getTerminator());
285           continue;
286         }
287       } else
288         continue;
289     }
290 
291     // The target does not have a single div/rem operation,
292     // and the rem is already in expanded form. Nothing to do.
293     if (!HasDivRemOp && E.isRemExpanded())
294       continue;
295 
296     if (HasDivRemOp) {
297       // The target has a single div/rem operation. Hoist the lower instruction
298       // to make the matched pair visible to the backend.
299       if (DivDominates)
300         RemInst->moveAfter(DivInst);
301       else
302         DivInst->moveAfter(RemInst);
303       NumHoisted++;
304     } else {
305       // The target does not have a single div/rem operation,
306       // and the rem is *not* in a already-expanded form.
307       // Decompose the remainder calculation as:
308       // X % Y --> X - ((X / Y) * Y).
309 
310       assert(!RemOriginallyWasInExpandedForm &&
311              "We should not be expanding if the rem was in expanded form to "
312              "begin with.");
313 
314       Value *X = E.getDividend();
315       Value *Y = E.getDivisor();
316       Instruction *Mul = BinaryOperator::CreateMul(DivInst, Y);
317       Instruction *Sub = BinaryOperator::CreateSub(X, Mul);
318 
319       // If the remainder dominates, then hoist the division up to that block:
320       //
321       // bb1:
322       //   %rem = srem %x, %y
323       // bb2:
324       //   %div = sdiv %x, %y
325       // -->
326       // bb1:
327       //   %div = sdiv %x, %y
328       //   %mul = mul %div, %y
329       //   %rem = sub %x, %mul
330       //
331       // If the division dominates, it's already in the right place. The mul+sub
332       // will be in a different block because we don't assume that they are
333       // cheap to speculatively execute:
334       //
335       // bb1:
336       //   %div = sdiv %x, %y
337       // bb2:
338       //   %rem = srem %x, %y
339       // -->
340       // bb1:
341       //   %div = sdiv %x, %y
342       // bb2:
343       //   %mul = mul %div, %y
344       //   %rem = sub %x, %mul
345       //
346       // If the div and rem are in the same block, we do the same transform,
347       // but any code movement would be within the same block.
348 
349       if (!DivDominates)
350         DivInst->moveBefore(RemInst);
351       Mul->insertAfter(RemInst);
352       Sub->insertAfter(Mul);
353 
354       // If X can be undef, X should be frozen first.
355       // For example, let's assume that Y = 1 & X = undef:
356       //   %div = sdiv undef, 1 // %div = undef
357       //   %rem = srem undef, 1 // %rem = 0
358       // =>
359       //   %div = sdiv undef, 1 // %div = undef
360       //   %mul = mul %div, 1   // %mul = undef
361       //   %rem = sub %x, %mul  // %rem = undef - undef = undef
362       // If X is not frozen, %rem becomes undef after transformation.
363       // TODO: We need a undef-specific checking function in ValueTracking
364       if (!isGuaranteedNotToBeUndefOrPoison(X, nullptr, DivInst, &DT)) {
365         auto *FrX = new FreezeInst(X, X->getName() + ".frozen", DivInst);
366         DivInst->setOperand(0, FrX);
367         Sub->setOperand(0, FrX);
368       }
369       // Same for Y. If X = 1 and Y = (undef | 1), %rem in src is either 1 or 0,
370       // but %rem in tgt can be one of many integer values.
371       if (!isGuaranteedNotToBeUndefOrPoison(Y, nullptr, DivInst, &DT)) {
372         auto *FrY = new FreezeInst(Y, Y->getName() + ".frozen", DivInst);
373         DivInst->setOperand(1, FrY);
374         Mul->setOperand(1, FrY);
375       }
376 
377       // Now kill the explicit remainder. We have replaced it with:
378       // (sub X, (mul (div X, Y), Y)
379       Sub->setName(RemInst->getName() + ".decomposed");
380       Instruction *OrigRemInst = RemInst;
381       // Update AssertingVH<> with new instruction so it doesn't assert.
382       RemInst = Sub;
383       // And replace the original instruction with the new one.
384       OrigRemInst->replaceAllUsesWith(Sub);
385       OrigRemInst->eraseFromParent();
386       NumDecomposed++;
387     }
388     Changed = true;
389   }
390 
391   return Changed;
392 }
393 
394 // Pass manager boilerplate below here.
395 
396 namespace {
397 struct DivRemPairsLegacyPass : public FunctionPass {
398   static char ID;
399   DivRemPairsLegacyPass() : FunctionPass(ID) {
400     initializeDivRemPairsLegacyPassPass(*PassRegistry::getPassRegistry());
401   }
402 
403   void getAnalysisUsage(AnalysisUsage &AU) const override {
404     AU.addRequired<DominatorTreeWrapperPass>();
405     AU.addRequired<TargetTransformInfoWrapperPass>();
406     AU.setPreservesCFG();
407     AU.addPreserved<DominatorTreeWrapperPass>();
408     AU.addPreserved<GlobalsAAWrapperPass>();
409     FunctionPass::getAnalysisUsage(AU);
410   }
411 
412   bool runOnFunction(Function &F) override {
413     if (skipFunction(F))
414       return false;
415     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
416     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
417     return optimizeDivRem(F, TTI, DT);
418   }
419 };
420 } // namespace
421 
422 char DivRemPairsLegacyPass::ID = 0;
423 INITIALIZE_PASS_BEGIN(DivRemPairsLegacyPass, "div-rem-pairs",
424                       "Hoist/decompose integer division and remainder", false,
425                       false)
426 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
427 INITIALIZE_PASS_END(DivRemPairsLegacyPass, "div-rem-pairs",
428                     "Hoist/decompose integer division and remainder", false,
429                     false)
430 FunctionPass *llvm::createDivRemPairsPass() {
431   return new DivRemPairsLegacyPass();
432 }
433 
434 PreservedAnalyses DivRemPairsPass::run(Function &F,
435                                        FunctionAnalysisManager &FAM) {
436   TargetTransformInfo &TTI = FAM.getResult<TargetIRAnalysis>(F);
437   DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
438   if (!optimizeDivRem(F, TTI, DT))
439     return PreservedAnalyses::all();
440   // TODO: This pass just hoists/replaces math ops - all analyses are preserved?
441   PreservedAnalyses PA;
442   PA.preserveSet<CFGAnalyses>();
443   return PA;
444 }
445